Trajectory recognition with a time-delay neural network

A time-delay neural network (TDNN) was used to classify object trajectories with varying positions, shapes, and directions and with different percentages of straight versus wavy segments. Dynamic recognition of these trajectories was accomplished so that as soon as a segment of sufficient length is available to the network, the type of target is identified by the output layer. The entire trajectory does not necessarily have to be input for recognition to occur, and overlapping trajectories can be recognized. Shift-invariance was attained by adjustment of bias weights during recall, as bias weights are a function of the amount of shift and the other weights after training on an initial trajectory is completed. Noise tolerance was tested on a series of sample trajectories. The network was found to be tolerant to up to 12% noise